ABSTRACTS
Forecasting Out-of-Hospital Cardiac Arrest Incidence Using Time Series ModelsAuthor: HyunHo Ryu | Professor | Chonnam national university hospital Associate Authors:
Background Accurate prediction of out-of-hospital cardiac arrest (OHCA) incidence is essential for optimizing emergency medical services and resource allocation. This study aimed to compare the performance of traditional time series models and deep learning algorithms in forecasting OHCA cases. Methods OHCA data from 2015 to 2020 were used to train the models, and data from 2021 were reserved for validation. We implemented multiple ARIMA and seasonal ARIMA (SARIMA) models with various parameter settings. For deep learning, an LSTM (Long Short-Term Memory) network was trained with varying window lengths and hyperparameters (epochs, batch size). Model performance was evaluated using Mean Absolute Percentage Error (MAPE) on both training and test sets. Results ARIMA models achieved consistent MAPE values ranging from 5.65% to 6.17%, demonstrating robust short-term forecasting with relatively low risk of overfitting. The LSTM model, depending on training duration and window size, yielded MAPE values between 5.24% and 5.77%. However, LSTM showed signs of overfitting when trained beyond 20 weeks of input data. The SARIMA model (ARIMA(2,0,2)(1,1,1)[12]) resulted in a higher MAPE of 11.53% on the test set. Conclusion While ARIMA models offer stable and interpretable predictions, LSTM can provide improved accuracy under optimized conditions. However, the risk of overfitting in long-term training should be carefully managed. These findings support the complementary use of both models in planning and preparedness for OHCA response systems.
|
